Efficient visibility algorithm for high-frequency time-series: application to fault diagnosis with graph convolutional network

Time series is a popular data type that is collected from various machines for fault diagnosis. Although most time-series models for fault diagnosis reflect local relations well, they cannot extract the global patterns that contain valuable information that can be used to recognize faults. To reflect the global structural information of a time series, many recent studies have used a graph constructed by visibility algorithms (VAs) that convert a time series into a graph. However, applying the VAs to high-frequency time series—which the machines typically generate—is challenging because the computational burden of the VAs increases with the length of a time series. Therefore, we propose a novel graph-based fault diagnosis framework for high-frequency time series. First, we propose an efficient VA (EVA) that extracts essential data points to characterize a time series and constructs a graph from a high-frequency time series. Not only do the EVAs convert a given time series faster into a graph than the VAs, but the resulting graphs also characterize the time-series structure with simplicity and clarity by selecting essential data points. Then, we adopt a graph convolutional network to analyze the resulting graphs and diagnose faults. We verified the characteristics of the EVAs and the fault diagnosis performance of the proposed framework using toy time series and public rotating machinery datasets, respectively. The results demonstrated that, compared to the VAs, the EVAs are efficient in terms of computational cost, and the proposed framework is effective for fault diagnosis.